Density-Sampled Gaussians (DeG) Representation

• DeG is a 3D representation where Gaussian centers are sampled from a learnable spatial density, enabling adaptive rendering and generative latent modeling.
• It combines an octree-based density factorization with Gaussian primitives to dynamically trade off rendering cost and visual fidelity.
• Differentiable adaptive density control replaces heuristic densification by integrating gradient-based optimization with a VAE framework for scalable, high-quality synthesis.

Searching arXiv for the DeG paper and closely related density-control work in 3D Gaussian representations. First, I’ll retrieve the main DeG paper and then relevant neighboring work on density control and densification in Gaussian splatting and feed-forward Gaussian models. Density-Sampled Gaussians (DeG) denotes a 3D representation in which Gaussian centers are not stored on a fixed voxel grid or array, but are instead modeled as samples from a learnable probability density function defined over an octree. In the 2026 formulation, DeG is designed to bridge the gap between adaptive rendering primitives and scalable generative modeling by jointly optimizing spatial density and Gaussian attributes under rendering supervision, using a differentiable render-loss contribution gradient as an analogue to the discrete densification and pruning heuristics of standard Gaussian Splatting. The same framework supports variable-resolution decoding from a single latent code by adjusting the sampling budget, and it is paired with a latent diffusion model equipped with VecSeq, a canonical re-indexing mechanism based on a deterministic 3D Sobol sequence (Yan et al., 8 May 2026).

1. Representation as a learnable spatial density

DeG represents the spatial support of Gaussian centroids as draws from a learnable PDF $q_\epsilon(x\mid Z)$ over $\mathbb R^3$, where $Z$ are the VAE latents. To make evaluation tractable at high resolution, this density is factorized via an $L$-level octree. If $x_{0:\ell}\in\{0,\dots,7\}$ denotes the child-index path from the root to level $\ell$, the density is written as

$$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$

Each factor is an 8-way categorical distribution produced by a small transformer cross-attending to $Z$. Sampling $P$ anchor points is performed by ancestral sampling down the octree, restricted to active branches. In practice, counts per cell are accumulated via systematic sampling to ensure exactly $P$ total samples, and each sampled leaf index is dequantized into a continuous 3D point by uniform jitter within its cell volume (Yan et al., 8 May 2026).

This construction makes the density itself an optimized object rather than an implicit consequence of repeated split/clone operations. A common misconception is that DeG merely renames heuristic densification in 3DGS. In the specific 2026 sense, DeG instead places the distribution of Gaussian centers under explicit probabilistic control through $\mathbb R^3$0, and uses that density as part of a generative latent-variable model. This suggests a conceptual shift from procedural density adjustment toward density as a learned, decoder-conditioned random field.

2. Anchor points, Gaussian primitives, and rendering

Once $\mathbb R^3$1 anchor points $\mathbb R^3$2 are drawn, each anchor spawns $\mathbb R^3$3 local Gaussians through learned offsets, giving a total of

$\mathbb R^3$4

primitives. The full set is $\mathbb R^3$5, where each primitive is parameterized by a mean $\mathbb R^3$6, covariance $\mathbb R^3$7, opacity $\mathbb R^3$8, and color $\mathbb R^3$9 or SH coefficients. In practice, $Z$0, while $Z$1, $Z$2, and $Z$3 are produced by a small transformer attending to $Z$4 and the anchor position (Yan et al., 8 May 2026).

Rendering is performed through differentiable Gaussian splatting, or “3DGS” rasterization. Along each camera ray $Z$5 for pixel $Z$6, transmittance and color are accumulated in depth order:

$Z$7

$Z$8

$Z$9

$L$0

followed by a background contribution $L$1. Here $L$2 is derived from $L$3, $L$4, and the ray intersection. The resulting image formation model is a standard volume integration of density and color (Yan et al., 8 May 2026).

Within this formulation, the representation is neither a pure point cloud nor a fixed-resolution radiance volume. It is an adaptive set of anisotropic volumetric primitives whose population is controlled by the learned octree density and whose local structure is expanded around sampled anchors.

3. Differentiable adaptive density control

The defining technical contribution of DeG is a differentiable mechanism for adaptive densification and pruning. Let $L$5 be the rendering loss computed from the Gaussian set generated by sampled anchors. Because anchors are sampled from $L$6, the training objective involves the expectation

$L$7

Using the score-function identity, the gradient with respect to the density parameters $L$8 becomes

$L$9

To reduce variance, DeG uses a leave-one-out “difference reward”

$x_{0:\ell}\in\{0,\dots,7\}$0

which measures how much the render loss increases if anchor $x_{0:\ell}\in\{0,\dots,7\}$1 is removed. The density-gradient estimator is then approximated by

$x_{0:\ell}\in\{0,\dots,7\}$2

The paper further states that $x_{0:\ell}\in\{0,\dots,7\}$3 extra renders are not performed. Instead, when $x_{0:\ell}\in\{0,\dots,7\}$4 contains an $x_{0:\ell}\in\{0,\dots,7\}$5 pixel term, the contribution of removing a primitive can be computed exactly inside the standard backward pass of 3DGS. If the rasterizer stores the transmittance $x_{0:\ell}\in\{0,\dots,7\}$6 and the “background color” behind primitive $x_{0:\ell}\in\{0,\dots,7\}$7, denoted $x_{0:\ell}\in\{0,\dots,7\}$8, then removing $x_{0:\ell}\in\{0,\dots,7\}$9 changes the pixel color by

$\ell$0

and the per-pixel $\ell$1 change is

$\ell$2

Summing $\ell$3 over pixels gives $\ell$4 for primitive $\ell$5, and primitives from the same anchor share the same $\ell$6 term (Yan et al., 8 May 2026).

This mechanism is presented as a fully differentiable analogue to discrete densification and pruning heuristics. The significance is methodological: density control is no longer imposed by external threshold logic, but is induced by render-loss contribution at the level of the learned sampling distribution itself.

4. Joint optimization and the DeG-VAE

DeG is trained as a VAE whose objective combines structural initialization, rendering supervision, regularization of Gaussian parameters, and a latent prior. The loss is written as

$\ell$7

Here $\ell$8 is the cross-entropy between ground-truth surface-point histograms and the octree density, $\ell$9 is the policy-gradient or difference-reward term, $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$0 is standard image reconstruction using $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$1, and $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$2 penalizes Gaussian volume, opacity, and cluster spread (Yan et al., 8 May 2026).

Training proceeds in three stages. Stage 1 uses only $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$3 to obtain a coarse hull. Stage 2 fixes the density and trains Gaussian attributes using $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$4. Stage 3 performs full joint training with all terms while randomizing $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$5 to encourage multi-resolution behavior. This staged procedure separates geometric support initialization, appearance learning, and fully coupled density-attribute optimization.

Within the broader literature on Gaussian scene representations, this training strategy situates DeG closer to latent-variable generative modeling than to per-scene optimization alone. It also makes density learning part of the decoder rather than a post hoc refinement pass.

5. Variable-resolution decoding and generative synthesis

A central property of DeG is variable-resolution decoding. At inference time, one chooses the anchor count $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$6, and because each anchor generates $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$7 local Gaussians, the total number of primitives is again $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$8. The paper states that this allows a smooth trade-off between rendering cost, which is $$q_\epsilon(x_{0:L}\mid Z) = \prod_{\ell=1}^L q_\epsilon(x_{0:\ell}\mid x_{0:\ell-1}, Z).$$9, and visual fidelity, and reports that PSNR/LPIPS improve continuously as $Z$0 increases (Yan et al., 8 May 2026).

To enable generative synthesis, the framework trains a latent diffusion model on DeG latents. Each 3D asset is encoded into an unordered set of $Z$1 latent tokens $Z$2. The paper identifies a critical challenge for diffusion on such latents: unordered set structure introduces $Z$3 equivalent permutations, which can significantly slow convergence. VecSeq addresses this by imposing a canonical spatial ordering using a fixed 3D Sobol sequence $Z$4. During preprocessing, FPS-sampled 3D points $Z$5 are matched to the Sobol sequence through an assignment $Z$6 minimizing

$Z$7

The reordered sequence is then augmented with a learned positional embedding of $Z$8, such as a 3D RoPE, before being processed by a conditional diffusion transformer, described as Flow Matching or Denoising Diffusion Transformer. The resulting sequence representation removes permutation ambiguity and is reported to yield much faster convergence and higher generation quality (Yan et al., 8 May 2026).

The abstract reports that the overall pipeline achieves state-of-the-art quality in single-image-to-3D generation. In that sense, DeG is not only a scene representation but also the decoding substrate of a generative model whose spatial adaptivity survives sampling-budget changes at inference time.

6. Relation to density-control literature and terminological scope

In the supplied literature, the label “Density-Sampled Gaussians” appears both as the name of a specific representation and as a broader descriptor for density-adaptive Gaussian distributions. The specific 2026 DeG formulation is best understood against a sequence of related approaches to Gaussian density control.

Paper	Core mechanism	Relation to DeG
"Generative Densification" (Nam et al., 2024)	Single-pass up-sampling, masking, decoding	Feed-forward densification of coarse Gaussians
"Frequency-Aware Density Control via Reparameterization for High-Quality Rendering of 3D Gaussian Splatting" (Zeng et al., 10 Mar 2025)	Density-scale mapping, dynamic densification, deletion	Density control via scale reparameterization
"Metropolis-Hastings Sampling for 3D Gaussian Reconstruction" (Kim et al., 15 Jun 2025)	Birth, relocation, voxel-penalized acceptance	Probabilistic alternative to heuristic cloning/splitting
"Gradient-Direction-Aware Density Control for 3D Gaussian Splatting" (Zhou et al., 12 Aug 2025)	GCR and nonlinear dynamic weighting	Gradient-direction-aware split/clone control

“Generative Densification” addresses generalized feed-forward Gaussian models rather than latent generative 3D modeling. It densifies Gaussians generated by a feed-forward backbone through a learned, single-pass up-sampling, masking, and decoding pipeline. On object-level reconstruction, the paper reports that baseline LaRa with 125 M parameters achieves PSNR $Z$9 on Gobjaverse, while GD-LaRa with 134 M parameters reaches PSNR $P$0, with SSIM improving from $P$1 and LPIPS from $P$2. On scene-level reconstruction, baseline MVSplat with 12 M parameters gives PSNR $P$3 on RE10K, while GD-MVSplat with 27.8 M parameters obtains PSNR $P$4, with SSIM $P$5 and LPIPS $P$6 (Nam et al., 2024).

FDS-GS establishes a direct relation between local Gaussian density and absolute scale through the reparameterization $P$7, and alternates dynamic-threshold densification with scale-based deletion. Its stated goal is a distribution in which low-frequency regions are represented by relatively large, low-density Gaussians and high-frequency regions by many small, high-density Gaussians. On the summarized averages over MipNeRF-360, the paper reports SSIM $P$8, PSNR $P$9, and LPIPS $P$0, compared with 3DGS at SSIM $P$1, PSNR $P$2, and LPIPS $P$3, while using 1–2 million fewer Gaussians than vanilla 3DGS (Zeng et al., 10 Mar 2025).

The Metropolis-Hastings approach reformulates densification and pruning as probabilistic sampling rather than threshold-based control. New Gaussians are inserted in under-represented, high-error regions using multi-view photometric and opacity scores, low-opacity Gaussians are relocated, and crowded voxels are discouraged by a sparsity prior and a voxel-factor in the acceptance rule. The paper states that this substantially reduces reliance on heuristics and uses fewer Gaussians while matching or modestly surpassing view-synthesis quality of state-of-the-art models (Kim et al., 15 Jun 2025).

GDAGS introduces the gradient coherence ratio

$P$4

and a nonlinear weighting

$P$5

to distinguish concordant and conflicting gradient directions during splitting and cloning. Across 13 scenes, the paper reports memory reductions of 40–50% while slightly boosting or matching reconstruction fidelity, and states 50% reduced memory consumption through optimized Gaussian utilization (Zhou et al., 12 Aug 2025).

Taken together, these works show that “density control” in Gaussian representations has evolved along several distinct lines: feature-space densification for generalized reconstruction, scale-constrained density adaptation, probabilistic sampling-based allocation, gradient-direction-aware split/clone logic, and, in DeG proper, direct learning of a density over space as part of a generative latent model. A plausible implication is that DeG occupies the point in this landscape where adaptive primitive allocation is made compatible with large-scale generative training, rather than being treated only as an optimization heuristic for reconstruction.